Google's next AI chip freezes Gemini into silicon to cut its power bill. The savings only count if they outrun demand.

The most important number in Google's new chip is not a speed. It is a ratio: tokens per watt. According to The Information, which broke the story last week, a processor Google is developing under the codename Frozen v2 is expected to serve the company's Gemini models somewhere between six and ten times more efficiently, per unit of power, than the custom silicon Google runs today. Six to ten times is a large number. Before deciding what it means, it is worth asking the question that decides whether any efficiency figure matters at all: efficient against what, and for how long.

Start with why the ratio, and not the raw speed, is the figure Google cares about. For most of the history of computing, the scarce thing was the chip. You wanted more performance, you bought more silicon, and the limit was how fast a fab could make it. That is no longer the binding constraint at the scale Google operates. The binding constraint is electricity — the megawatts a data centre can actually draw from the grid it sits on. And megawatts do not arrive on a chip-industry schedule. They arrive on a utility's schedule, which means interconnection queues measured in years, transformer lead times measured in years, and, increasingly, state regulators writing rules that treat a large data centre the way they would treat a new factory or a small city. When the power is capped, you cannot add compute by adding machines. You can only add compute by doing more with the power you already have. That is what tokens per watt measures, and it is why it has quietly become the number the whole industry is now organised around.

What it means to freeze a model into a chip

A conventional AI accelerator — Google's own TPU, or an Nvidia GPU — is a general machine. It is built to run whatever model you load onto it: this architecture today, a different one next year, someone else's entirely the year after. That flexibility is worth a great deal, and you pay for it in energy. A general chip spends a large share of its power not on the arithmetic that produces an answer but on moving numbers around — shuttling weights and activations between memory and the processor, over and over, because it cannot assume in advance what shape the computation will take.

Frozen v2, as reported, refuses that generality. It hardwires the shape of Gemini — the arrangement of layers and operations that defines the model, its blueprint — directly into the silicon. Because the chip already knows the structure of the computation it will run, it can carry enough memory on board to hold the model without constantly reaching off-chip, and it can fuse operations that a general processor would perform one at a time. Less data movement, fewer wasted steps, and the energy that a flexible chip spends staying flexible is simply not spent. That is where a factor of six to ten comes from. It is not a faster transistor. It is the removal of work the chip no longer has to do.

The trade is stated plainly in the design itself, and it is the whole story: you buy those joules with flexibility. A chip cast around one model's blueprint is very good at that model and useless at anything shaped differently. Which raises the obvious risk, and Google has clearly already run into it once.

They tried freezing the weights. It broke.

The idea, by the reporting, originated with Jeff Dean, Google's chief scientist, and there was an earlier version. The first Frozen design tried to go further — to bake not just the model's architecture but its actual trained parameters, its weights, into the hardware. On paper that is the most efficient thing you could possibly build: the entire model, numbers and all, frozen into the circuit. In practice it was abandoned, because a chip with the weights cast into it is locked to one specific version of Gemini. The moment the model is retrained — which, for a frontier model, is roughly always — the chip is obsolete. You would be manufacturing silicon with the shelf life of a software release.

Frozen v2 is the lesson learned. It freezes only the blueprint, not the parameters, so new weights can still be loaded onto it. The architecture stays put; the model it holds can keep learning. That is a genuinely clever hedge, and it is worth crediting as engineering. But it does not remove the underlying bet. It narrows it. The wager is no longer that a single trained model will stay useful. It is that the shape of the model — the architecture itself — will hold still long enough to be worth casting in silicon and running until 2028, the year Google reportedly intends to deploy the thing, and for years after that. This is the number that decides whether the chip pays for itself, and it is not a number about power at all. It is a bet about how fast the field stops changing.

The wager is that the shape of the model holds still long enough to be worth casting in silicon. That is not a bet about power. It is a bet about how fast the field stops moving.

On the current evidence, I would price that bet carefully. The same week the Frozen v2 report surfaced, Google's DeepMind team said it had begun its most ambitious pre-training run yet, for Gemini 4. Architectures are still moving between numbered generations, not settling into a standard. And underneath the specific models, the transformer — the basic structure that every large model has shared for most of a decade — is itself under active challenge from researchers pushing state-space models, linear-attention schemes, and other designs meant to be cheaper than the thing Frozen v2 would freeze. None of that is a reason the chip won't work. It is a reason to read '2028' as the real risk in the plan. A custom chip is only cheap if you run enormous volumes of the workload it was built for. Freeze the wrong blueprint a year before the field moves, and the most efficient chip in the building becomes a very expensive way to run a model nobody uses any more.

The efficiency is real. It is also a treadmill.

Suppose the bet lands — the architecture holds, the chip ships in 2028, and it delivers its six-to-tenfold gain. Here is the part the headline number quietly omits: efficiency per token is not the same as total power, and only one of those two things shows up on the grid. Total power is efficiency multiplied by volume. Cut the energy per token by a factor of ten, and the power a data centre draws falls only if the number of tokens does not rise by more than a factor of ten. Every signal says it will rise by far more than that.

Consider what Google itself just did to its own token volume. Earlier this month the company made a Gemini model the thing that answers every search query — not an opt-in feature, but the default behaviour of the search box, worldwide. A single language model now composes a response to a meaningful share of the questions the planet asks in a day. That is not a marginal increase in inference load. It is a step change in it, and it is exactly the kind of demand that a cheaper cost-per-token invites. This is the oldest pattern in energy economics, old enough to have a name — make a resource cheaper to use and people use so much more of it that total consumption climbs. A more efficient engine did not reduce the world's demand for fuel; it expanded the number of things worth burning fuel to do. There is no reason to expect tokens to behave differently, and Google's own product decisions are actively ensuring they won't.

So price the claim honestly. Frozen v2, if it works, lowers the cost of a token and the marginal power a token draws. Both are real and both are worth having. What it does not do is lower the total electricity Google's data centres pull from the grid, because the company is growing its token volume faster than any single chip can shrink the power behind each one. The efficiency is not an exit from the grid wall. It is a way of arriving at that wall more slowly — of fitting more model behind a fixed number of megawatts — which is worth doing, and is a different thing from escaping the constraint.

Who this is really for, and where the logic travels

It helps to be precise about what Frozen v2 is and isn't. It is reportedly an internal chip, for Google's own data centres, in smaller volume than its main TPU line, aimed squarely at cutting the cost of running Gemini — which is to say, at margin. Every watt Google saves serving an answer is a cent it does not have to charge, or a cent it keeps, at a moment when it is competing on the price of inference against OpenAI and Anthropic and pouring a reported one hundred and ninety billion dollars into capital spending this year. The chip is a cost-curve move dressed, in the coverage, as a breakthrough. Alphabet reports quarterly earnings the same week the design leaked; the market read the report as good news and marked the shares up. That reaction is about the cost of compute, not the physics of it, and it is correct on its own terms.

The reason this matters beyond Google's balance sheet is that the logic does not stay in Oregon or Iowa. The megawatts a hyperscaler saves in a well-supplied American grid do not appear on an invoice in Lagos or Jakarta. But the metric does travel, and it is the one that decides whether this technology can physically exist in the places where power is scarcest and dearest. In a grid that cannot spare the load, tokens per watt is not an accounting nicety — it is the difference between a model that can be served locally and one that cannot be served at all. The unglamorous discipline of squeezing more computation out of each joule, of designing the waste out of the hardware, is exactly the kind of engineering that determines where compute is allowed to live. It is worth watching for that reason, whatever happens to Google's margins.

For the frozen chip specifically, the case rests on a short list of things all being true at once:

• The efficiency claim survives contact with a real workload — six to ten times per watt is, for now, an internal expectation reported at second hand, not a measured, deployed result.
• The Gemini architecture holds its shape from now through 2028 and well beyond, long enough for a fixed-function chip to earn back what it cost to design and fabricate.
• The per-watt gains keep compounding faster than Google's own token volume grows — the condition under which efficiency actually relieves the grid rather than merely feeding more demand into it.

The first is plausible; hardwiring a known computation is a well-understood way to save energy, and the size of the claimed gain is consistent with cutting out data movement. The second is a real bet against a field that is still visibly in motion. The third is the one almost no one mentions, and it is the one that decides whether any of this changes the number that actually constrains the industry. Frozen v2 is a serious, sensible piece of engineering aimed at the right variable. But the variable it improves is efficiency, and efficiency, in a business growing this fast, has a way of being spent as quickly as it is saved. The grid wall does not move because a chip got frugal. It moves when the megawatts arrive — and the megawatts are still stuck in the queue.